Igor's Lab believes there's something wrong with the scheduler that can't be fixed with drivers. It's been covered by Linus and Moore's Law is Dead.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Geekbench score for the Intel i9-13900K is out and it’s not looking pretty for Apple …
- Thread starter Retskrad
- Start date
- Sort by reaction score
I guess we will know if they send out another die revision for A380 that is different than what is in folks hands now.
I'd very much like Intel to succeed with Arc. Having an extra competitor in the GPU space would be great. I'm just not sure that they have the institutional fortitude for it. Larrabee followed a similar pattern, with expectations of a gaming card, but it ended up being shoved into the data center. The spiritual forefather of Alchemist, the i740, never made it to version two. Hopefully, Arc breaks the curse.
Which failed because folks wanted backwards compatibility (which Titanium doesn't offer).
I mean they are committed to 3 generations at least, let’s see what happens. I think the Intel Arc GPUs will be quite helpful for Intel in the laptop space especially the integrated and cheap gaming laptops space
If AMD didn’t exist at the time, if Intel REALLY had the market control that would mean that customers would continue to stick with them as they removed the backwards compatible stuff and made the replacement more performant, Intel would be in a better place today with a more streamlined simplified and efficient instruction set. As it is, they were forced to backpedal and here we are, at a point where the chip is so complex, it needs assistance from the OS scheduler to hit it’s full performance.
Eh, I think some of that is due to Intel not making consumer IA64 chips like AMD made consumer AMD64 ones. I think Intel believed they could get the server space to move, then maybe the consumer side. It didn't work out that way (clearly).
Its gimped. If you think a mobile gpu from Apple plays games perfectly fine, then you clearly have a very loose definition of what a good performing GPU is.
Until we have a native AAA title on macOS with Metal 3 we cannot judge gaming pref on any M1. Resident Evil Village should give us a GREAT example on how Apple's GPU's perform when they are not hobbled back by old OpenGL or Rosseta translation.
Apple's GPUs are good. They on par with RDNA 2 iGPUs.
Right, what I said was, “If AMD didn’t exist at the time”. And, if AMD didn’t exist, they wouldn’t have been making consumer AMD64 anything. Without AMD, the industry would have been pissed with Intel, sure, but there literally wouldn’t have been any other way forward and they’d just make the best of Intel’s offerings. Just like the industry made the best of the Skylake situation.
Apples GPUs are capable of very good performance for their power envelope. Expecting them to perform on par with GPUs with several times their power draw is clearly unreasonable (particularly on code targeting the PC offerings).
Conversely, ignoring the power side of the equation is also unreasonable since Apple clearly designs their SoCs to allow ergonomics and form factors they favour.
We’ll see where Apple goes from here, but it would surprise me greatly if they were to drastically change their design ethos to line up with desktop PCs. In my book, this is a good thing.
On one hand it is cool that the M2 can out perform the 680M, on the other since it has more ALU's it seems like it should outperform the AMD part right?
M2 gets ~ 35FPS at 1080p medium in BG3 and scores over 6k in 3Dmark Wild Life. That is compatible with Nvidia's current lower-power 25-30W dedicated GPUs. Of course it's far from enthusiast gaming level but it's absolutely sufficient for an everyday user. As to the rest, refer to the excellent post by @EntropyQ3 above.
we can't do a proper comparasion with both these chips until M2 has a game that is fully optimised with the API.
That's why RE: Village is going to be a good benchmark because it will made using Metal 3. So we can finally see just how good the GPU in the M2 is compared to the 680M.
Synthetic benchmarks are a decent proxy of the potential performance. In 3DMark, M2 is faster than the 680M by at least 20% and is positioned between the Max-Q and regular laptop variants of GTX 1650.
Real game results are obviously going to the the decisive test, but just because a game has been ported to Metal 3 doesn't mean that the port has been done well. One can hope of course. At any rate, I intend to purchase these games even if it's not really my cup of tea just to send the signal to the market
It's interesting just how close that was to happening. According to Cliff:
So, most of the design team left AMD, only 15-20 were left, and they had a vote whether to stay. In some alternate timeline, we're all using Itanium, like it or not.
Maybe we would have been better off if x86 had died back then.
Perhaps, we'll never really know. Cliff says that Intel "let the HP PA-RISC guys make a science project" and the end result was Itanium. Given his time at Intel's competitors, I understand why he's not a fan of team blue, but we'll ultimately never know. The market simply wouldn't accept non-x86 mass market CPUs at that time because AMD provided an alternative, and it's only now that we're getting some measure of competition with a different ISA. It's reminiscent of the 80s, when there were multiple competing designs.
igor’s Lab is reaching for a root cause . There is a retest with updated drivers .
Intel Arc A380 re-benchmarked: New drivers and the draw calls - API overhead on different platforms | igor´sLAB
New driver, new luck? I gave Intel’s Arc A380 with new beta drivers another go and also looked at the draw calls under DX11, DX12 and Vulkan. Old wine and maturing? Yep, but I’ll get to that in a…
www.igorslab.de
Intel doesn’t have a mega framerate gaming card here . Reviews highly skewed around the card‘s worth almost totally hindged on being a Nividia ’killer’ card are going to drift into pointing at hardware being ‘broken’ cause(s) .
much of the larger Xe core history was aimed more at workstation / server GPGPU compute than more maximum frame rate Monitor driving .just as much as a ‘hardware problem’ the results look like Intel is trying to goose the max frame rates too high and crappy edge cases pop up . You can frame that in the context of ‘ if hardware was faster you could flog it harder without glitches’ , but that also can be coach as match the drivers to what you have , not what you wished you had.
Intel has probably made a mistake chasing after gaming fanboy benchmarks. DirectX11 isn’t a focus area , but is soaking up resources. So is chasing after funky game engine stack 27 with quirks . If Intel had gone after mainstream apps , content creation with overlap with their leading Video de/encode abilities , mandatory resiable BAR, and 60-80 Hz frame rates they would probably be in better shape. OpenCL , OneAPI stuff that are aiming the Xe-HPC card at is probably a better fit with the shared Xe Core linage.
The cryptocaraze probably also plays a role here . Both in jumbling Intels’s foscus ( which not so great at anyway ) and worrying about hash rate limiters . [ if gpu draw call rate is slow is a hash calls rate going to vast multiples faster ? ]
what Intel really needed/needs was a good enough performance , cost effective cards for system vendors to include in systems for regular folks , not folks focused on games . That would not have ‘killed’ Nvidia but would give them a foothold closer to the iGPU market they dominated already. make it work and then make it fast In gen 2 and 3 . The volume would have been lower but the first gen should not need to be a high volume seller to be successful.
if look at the line up ( without he hype train from sales ) that is mainly what the hardware are looks like( low to midrange affordable option) . Unfortunately, some part of Intel Sales marketing put the gaming abilities on a hype train and some middle and upper management drank the company kool aid ( and/or got talked into making gaming reviewers happy to sell cards ) . Maybe also foolishly thinking the GPU supply scarcity was going to allow them time to chase after too many directions at same time and folks would just buy because it was there at close to MSRP . If so , That was huge blunder .
P.S. if there is a hardware problem I suspect at least part of that was Intel trying to put DisplayPort 2.0 on these . That was probably ‘ a bridge to far’ given scope of the work and newness to the territory. If there is firmware updates for display controllers that might be an contributing issue fix that could come with more time . ( Apple is not being very sgreesive there at all . M2 still has same video streams out limits Save complexity on sub component X so can add more polish to sub component Y . )
My understanding was that, byte for byte, Optane drew more power than SSD. It could handle direct access/modification of individual bytes, unlike NAND storage, which has to transfer blocks of bytes at a time. Hence, in theory, the durability of individual cells could be greater, since they do not necessarily have to be involved in block writes.
However, data storage is generally handled in blocks rather than tiny fragments, and there really are practical advantages to doing it that way (like managing stored versions and keeping them consistent). In the end, it looks a bit like Optane was providing a solution to a problem that may not currently call for that approach, and possibly never will (I can easily see high-density NVMe embedded inside an SoC to be used as program code source, but that is memory that would be infrequently written).
Perhaps the biggest failure was that it never migrated to mobile platforms, where it might make a lot of sense, in part because the power budget of NAND + DRAM is still enough better than Optane to make the switch not worthwhile. But ultimately Intel and Micron were unable to make it pay for itself.
Larrabee was actually a supercomputer chip first and foremost, something intended to compete with the then-new phenomenon of GPGPU in supercomputing. The ability to repurpose the chip as a GPU that would rasterize graphics was kind of a side project / goal.
For the full story from an insider, go here and click the "Why didn't Larrabee fail?" link in the sidebar on the right.
TomF's Tech Blog - It's only pretending to be a wiki.
tomforsyth1000.github.io
I've never worked for Intel or its competitors and I'm more than willing to just say it: the Itanium ISA was extremely stinky dog poop. It was so bad it just doesn't seem likely it could have taken over from x86 on technical merits, even though x86 itself is not an ideal ISA.
It's not clear when Intel would have tried to do that either, as they positioned Itanium as their answer to RISC workstations and servers, not something for mass market personal computers.
"Science project" is apt. It's a mashup of far-out ideas. Each is interesting in the academic sense, as in it might have made sense for a researcher to write a paper about the idea and figure out whether it had any merit. Instead of doing something like that, Intel just jammed a bunch of these ideas into their ISA of the future and pushed it into production without doing the homework to rigorously check whether any of it was likely to work out.